Here's the mistake almost every analyst makes with AI. They ask which model is best. ChatGPT, Gemini, Claude, Perplexity. They pick one and run everything through it. The failure mode is quiet. You hand one tool a job it was never built for, it gives you a confident answer, and you never check it. On a research desk that does not hold up.
So flip the question. The desk is not one job. It is five. Research and synthesis. Reading filings and calls. Writing the memo. Verifying a fact. Monitoring your work against your own sources. One job, one tool, then you verify. That frame comes from a Jeff Su walkthrough of his stack, and it maps cleanly onto how we actually work.
Job one. Research and synthesis. This goes to a general purpose chatbot, and the thing you want is instruction following. The strongest one takes a long checklist and does not drop steps. Here's why that matters. Your research prompt is a checklist. Compare this quarter to last. Pull the revenue drivers. Pull margins. Pull guidance. Pull the risk factor changes. Flag any shift in management language. A model that quietly skips three of those is worse than useless, because it looks finished and it is not. So structured research and your first synthesis go to the model that actually follows orders.
Job two. Reading mixed material. This goes to a multimodal model, one that can take video, audio, images, and text at once. Picture your real inputs the morning after an earnings event. A recording of the call. The slide deck. A photo of an exhibit you screenshotted off the print. Most tools can only read the text. One model takes all of it together and tells you what was discussed and what changed. For an analyst that is the difference between transcribing for an hour and reading a synthesis in a few minutes. So ingesting calls, decks, and exhibits goes to the multimodal model.
Job three. The last mile. The draft you actually send and the small script you keep punting on. This goes to a draft quality model, the one where the first attempt is closest to done. Two pieces matter here. First, voice. Feed it your past memos and it matches your house style, so the draft reads like your desk, not like a robot. Second, working code. Jeff needed to bulk export data, asked for a script, and it ran on the first try. He says he cannot write code at all. For us that is the data pull you always hand to someone else. Cleaning a CSV. Reshaping a comp table. Charting a series. So polishing the memo and writing the quick pull script go to the draft quality model.
Job four. The fast fact. This goes to a search specialist tool. There's a clean line here. Chatbots are built for reasoning. Search tools are built for fetching. So you do not ask a search tool to think. You ask it for one specific, current fact, and you want it now. Was guidance reiterated or cut. What is the current share count. When is the next print. Think of it as the search scalpel. It checks the thing the reasoning model just told you. That's the move. The chatbot drafts the thesis. The search tool verifies the facts inside it.
Job five. This is the one I care about most. Monitoring your work against your own sources. This goes to a source grounded tool, one that answers only from the documents you upload. It has no outside knowledge to invent from. Here's the workflow. Before Jeff publishes, he uploads his draft plus the source material and asks it to flag any claim the sources do not support. Translate that straight to a memo. Upload your draft and the actual filing. Ask one question. Does this memo make any claim the filing does not support. That catches the quiet errors. The number that drifted a decimal. The guidance you remembered slightly wrong. So checking your work against your own primary sources goes to the source grounded tool.
Now the part most tool videos skip. The heads up. A source grounded tool is only as good as the sources you feed it. Bad source in, confident wrong answer out. In finance that is the whole game. And notice what none of these tools did. None of them decided anything. They draft, they read, they fetch, they check. You own the assumptions. You own the source list. You own the call.
So the rule does not move. The tool builds. You verify against the primary filing. No source, no claim. None of this is buy, sell, or hold advice, and the tool names are workflow examples, not endorsements. AI structures the research process. It does not pick the stock, and it does not replace the filing. Treat every output as a draft until you have traced the number back to the primary source yourself. The goal is to be better prepared, not more confidently wrong. Map your five jobs to five tools, and verify every one.